NeurIPS 2023 Papers — Page 21
Conference on Neural Information Processing Systems · 3218 papers
Offline RL with Discrete Proxy Representations for Generalizability in POMDPs
Pengjie Gu (Nanyang Technological University), Bo An (Nanyang Technological University)
Autonomous DrivingRecurrent Neural NetworkReinforcement LearningAuto EncoderTabular
🎯 What it does: The ORDER framework is proposed to enhance the generalization performance in different partially observable POMDP environments through discrete agent representation in offline reinforcement learning.
OKRidge: Scalable Optimal k-Sparse Ridge Regression
Jiachang Liu (Duke University), Cynthia Rudin (Duke University)
OptimizationTabular
🎯 What it does: This paper presents OKRidge, a branch-and-bound algorithm that utilizes a new saddle point lower bound and ADMM to solve k-sparse ridge regression in a provably optimal manner.
On Calibrating Diffusion Probabilistic Models
Tianyu Pang (Sea AI Lab), Zhijie Deng (Tsinghua University)
GenerationDiffusion modelImageStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: This paper proposes a technique that only requires a single calibration to improve the model score matching loss and likelihood lower bound for existing diffusion probability models (DPM).
On Certified Generalization in Structured Prediction
Bastian Boll (Heidelberg University), Christoph Schnoerr
🎯 What it does: A PAC-Bayesian generalization risk upper bound for structured prediction is proposed.
On Class Distributions Induced by Nearest Neighbor Graphs for Node Classification of Tabular Data
Federico Errica (NEC Laboratories Europe)
ClassificationGraph Neural NetworkTabular
🎯 What it does: This paper studies the use of k-NN graphs for node classification in tabular data without original graph structure and constructs a theoretical framework to evaluate its impact on deep graph networks.
On Computing Pairwise Statistics with Local Differential Privacy
Badih Ghazi (Google Research), Adam Sealfon (Google Research)
Safty and Privacy
🎯 What it does: This study investigates the problem of computing pairwise statistics under local differential privacy, proposing several novel algorithms that can effectively compute important metrics such as Kendall's τ coefficient, area under the curve, and Gini mean difference.
On Convergence of Polynomial Approximations to the Gaussian Mixture Entropy
Caleb Dahlke (University of Arizona), Jason Pacheco (University of Arizona)
🎯 What it does: This paper studies a polynomial approximation method for the entropy of Gaussian mixture models and provides a convergence analysis.
On Differentially Private Sampling from Gaussian and Product Distributions
Badih Ghazi (Google Research), Pasin Manurangsi (Google Research)
Safty and Privacy
🎯 What it does: This paper studies algorithms for generating approximate samples from unknown high-dimensional Gaussian distributions and binary product distributions under differential privacy (DP) constraints, providing upper and lower bounds on sample complexity under several different assumptions;
On Dynamic Programming Decompositions of Static Risk Measures in Markov Decision Processes
Jia Lin Hau (University of New Hampshire), Marek Petrik (University of New Hampshire)
OptimizationReinforcement LearningFinance Related
🎯 What it does: The study conducts dynamic programming decomposition for static risk measures (CVaR, EVaR, VaR) in Markov Decision Processes (MDP) and reveals the inherent suboptimality issues of commonly used CVaR and EVaR decomposition methods in policy optimization; it also proposes and proves a correct dynamic programming decomposition applicable to VaR.
On Evaluating Adversarial Robustness of Large Vision-Language Models
Yunqing Zhao (Singapore University of Technology and Design), Min Lin (Sea AI Lab)
Adversarial AttackTransformerVision Language ModelDiffusion modelImageTextMultimodality
🎯 What it does: This study investigates the robustness of large-scale visual language models (VLM) in black-box and targeted attack scenarios, and proposes a hybrid attack method based on transfer and query, successfully inducing various open-source VLMs to generate predetermined target texts.
On Generalization Bounds for Projective Clustering
Maria Sofia Bucarelli (Sapienza University of Rome), Mads Toftrup
Tabular
🎯 What it does: This paper studies the generalization error upper bounds of point center clustering (k-median, k-means) and subspace clustering (especially item clustering). It provides nearly optimal upper bounds when the sample size n, the number of clusters k, the subspace dimension j, and the error index z are constants, and proves that the corresponding lower bounds match previous work.
On Imitation in Mean-field Games
Giorgia Ramponi (ETH AI Center), Matthieu Geist (Google DeepMind)
Reinforcement LearningAgentic AI
🎯 What it does: This paper studies the problem of imitation learning in Mean-field Games, proposing the Nash imitation gap metric and analyzing its theoretical properties.
On kernel-based statistical learning theory in the mean field limit
Christian Fiedler (RWTH Aachen University), Sebastian Trimpe (RWTH Aachen University)
🎯 What it does: This paper studies the theoretical aspects of kernel methods in the mean field limit as the number of variables approaches infinity, including the limiting properties of kernels and their RKHS, approximation properties, and the convergence of support vector machines.
On Learning Latent Models with Multi-Instance Weak Supervision
Kaifu Wang (University of Pennsylvania), Dan Roth (University of Pennsylvania)
ClassificationImage
🎯 What it does: This paper proposes a multi-instance partial label learning (multi-instance PLL) framework, studying learnability and error bounds under unknown or known deterministic transformation functions.
On Learning Necessary and Sufficient Causal Graphs
Hengrui Cai (University of California), Rui Song (North Carolina State University)
OptimizationGraph Neural NetworkGraphTabular
🎯 What it does: A new necessary and sufficient causal graph learning algorithm, NSCSL, is proposed, which can simultaneously filter causal features and learn causal structures in high-dimensional data.
On Masked Pre-training and the Marginal Likelihood
Pablo Moreno-Muñoz (Technical University of Denmark), Søren Hauberg (Technical University of Denmark)
TransformerAuto EncoderImageText
🎯 What it does: Through theoretical derivation and empirical validation, it is proven that Masked Pre-Training (MPT) essentially randomizes the maximization of the model's marginal likelihood (LML), revealing the equivalence between its cumulative loss and LML;
On Measuring Fairness in Generative Models
Christopher T.H Teo, Ngai-man Cheung
GenerationData SynthesisDiffusion modelGenerative Adversarial NetworkImage
🎯 What it does: The study measures the fairness of generative models and finds that existing methods still produce significant errors even when using high-accuracy attribute classifiers, proposing a fairness measurement framework based on classifier error called CLEAM.
On permutation symmetries in Bayesian neural network posteriors: a variational perspective
Simone Rossi (Stellantis), Thomas Hannagan (Stellantis)
OptimizationConvolutional Neural NetworkImage
🎯 What it does: This study investigates and verifies the property that there exists a linear connectivity between Bayesian Neural Network (BNN) solutions obtained through Variational Inference (VI), which can be achieved with almost no loss by aligning through weight permutation.
On Private and Robust Bandits
Yulian Wu (King Abdullah University of Science and Technology), Di Wang (King Abdullah University of Science and Technology)
OptimizationSafty and Privacy
🎯 What it does: For the multi-armed bandit (MAB) problem under the Huber contamination model with heavy-tailed rewards and differential privacy requirements, a unified meta-algorithm is proposed that achieves private and robust mean estimation through truncation and the Laplace mechanism, thereby achieving approximate optimal lower and upper bounds on the regret rate.
On Proper Learnability between Average- and Worst-case Robustness
Vinod Raman (University of Michigan), Ambuj Tewari (University of Michigan)
OptimizationAdversarial Attack
🎯 What it does: The study achieves learnability of VC classes by relaxing the robust loss function in adversarial robustness learning.
On quantum backpropagation, information reuse, and cheating measurement collapse
Amira Abbas (Google Quantum AI), Jarrod Ryan McClean (Google Quantum AI)
Reinforcement LearningPhysics Related
🎯 What it does: This paper explores whether the gradient estimation of parameterized quantum models can achieve efficient training like classical backpropagation, and proposes a quantum backpropagation algorithm that utilizes multiple copies of quantum states and shadow measurements.
On Robust Streaming for Learning with Experts: Algorithms and Lower Bounds
David Woodruff, Samson Zhou (Texas A&M University)
🎯 What it does: The study focuses on online expert learning algorithms for adaptive inputs under limited memory, providing both deterministic and stochastic robust implementations.
On Sample-Efficient Offline Reinforcement Learning: Data Diversity, Posterior Sampling and Beyond
Thanh Nguyen-Tang (Johns Hopkins University), Raman Arora (Johns Hopkins University)
Reinforcement Learning
🎯 What it does: This paper proposes a new measure of data diversity and uses this measure to unify and compare three types of offline reinforcement learning algorithms (version space, regularized optimization, and posterior sampling), proving that they can achieve the same order of sample efficiency under standard assumptions.
On Separate Normalization in Self-supervised Transformers
Xiaohui Chen (Tufts University), Liping Liu
Representation LearningTransformerImageTextGraph
🎯 What it does: This paper proposes SepNorm, a technique for separating the normalization of the [CLS] vector from that of regular token vectors in self-supervised Transformers;
On Single-Index Models beyond Gaussian Data
Aaron Zweig (New York University), Joan Bruna (New York University)
OptimizationTabular
🎯 What it does: The study investigates whether the single-index model (nonlinear labels after one-dimensional projection) can effectively recover the hidden direction θ∗ using stochastic gradient descent (SGD) under non-Gaussian data distributions.
On skip connections and normalisation layers in deep optimisation
Lachlan Ewen MacDonald, Simon Lucey (Australian Institute for Machine Learning)
OptimizationConvolutional Neural NetworkImage
🎯 What it does: This paper constructs a general multi-layer parameterized system (MPS) theoretical framework, systematically analyzing the effects of batch normalization, weight normalization, and residual skip connections on network loss curvature and regularization. Within this framework, it is proven that gradient descent can converge to a global optimum in networks containing these structures (even if the global optimum is located at infinity). Additionally, the authors reveal the causal mechanism by which residual connections enhance the Jacobian singular value distribution, thereby accelerating training through experiments on singular value distribution.
On Slicing Optimality for Mutual Information
Ammar Fayad (Massachusetts Institute of Technology), Majd Ibrahim (Higher Institute for Applied Sciences and Technology)
OptimizationReinforcement LearningGenerative Adversarial NetworkImageSequential
🎯 What it does: This paper proposes an optimal sliced mutual information (SI*) for measuring the dependence of high-dimensional random variables;
On Sparse Modern Hopfield Model
Jerry Yao-Chieh Hu (Northwestern University), Han Liu (Northwestern University)
OptimizationRecurrent Neural NetworkContrastive LearningImage
🎯 What it does: A sparse modern Hopfield model is proposed, establishing a theoretical equivalence with the sparse attention mechanism;
On student-teacher deviations in distillation: does it pay to disobey?
Vaishnavh Nagarajan (Google Research), Sanjiv Kumar (Google Research)
ClassificationKnowledge DistillationConvolutional Neural NetworkImageText
🎯 What it does: This study investigates the systematic bias in the predicted probabilities of students and teachers in knowledge distillation and explains why students can achieve better generalization while deviating from the teacher.
On the Ability of Graph Neural Networks to Model Interactions Between Vertices
Noam Razin (Tel Aviv University), Nadav Cohen (Tel Aviv University)
Graph Neural NetworkGraph
🎯 What it does: The theoretical analysis of the ability of Graph Neural Networks (GNN) to capture interactions between vertices is conducted, and based on this, a sparsification algorithm called WIS (Walk Index-based Sparsification) is proposed, aiming to maintain the interaction modeling capability of GNN when edges are deleted.
On the Adversarial Robustness of Out-of-distribution Generalization Models
Xin Zou (Wuhan University), Weiwei Liu (Wuhan University)
Domain AdaptationAdversarial AttackImageBenchmark
🎯 What it does: This paper studies the adversarial robustness in out-of-distribution (OOD) models, finding that existing OOD methods are vulnerable to adversarial attacks. It subsequently proposes two theory-based algorithms (AT and RDANN) to enhance OOD adversarial robustness.
On the Asymptotic Learning Curves of Kernel Ridge Regression under Power-law Decay
Yicheng Li (Tsinghua University), Qian Lin (Beijing Academy of Artificial Intelligence)
Tabular
🎯 What it does: This paper studies the asymptotic learning curve of kernel ridge regression under power-law decay conditions, providing a complete characterization of the learning curve and exploring the interactions between the regularization parameter, source conditions, and noise.
On the choice of Perception Loss Function for Learned Video Compression
Sadaf Salehkalaibar (University of Toronto), Ashish J Khisti
CompressionFlow-based ModelVideo
🎯 What it does: This study investigates the impact of two perceptual loss functions on reconstruction quality in causal, low-latency video compression, and proposes the universality of MMSE representation.
On the Complexity of Differentially Private Best-Arm Identification with Fixed Confidence
Achraf Azize (University of Lille), Debabrota Basu (University of Lille)
OptimizationSafty and PrivacyTabular
🎯 What it does: This paper studies the best arm identification problem under fixed confidence (FC-BAI), proposes a sample complexity lower bound under global ε-differential privacy (DP) constraints, and designs a new algorithm AdaP-TT that achieves the same order as the lower bound at a high privacy level.
On the Connection between Pre-training Data Diversity and Fine-tuning Robustness
Vivek Ramanujan (University of Washington), Ludwig Schmidt (University of Washington)
ClassificationData-Centric LearningConvolutional Neural NetworkSupervised Fine-TuningImage
🎯 What it does: This study investigates the impact of different pre-training data distribution attributes (data volume, label granularity, label semantics, image diversity, data sources) on the robustness of fine-tuned models under iWildCam-WILDS, and proposes an 'effective robustness' evaluation framework.
On the Consistency of Maximum Likelihood Estimation of Probabilistic Principal Component Analysis
Arghya Datta (Université de Montréal), Sayak Chakrabarty (Northwestern University)
🎯 What it does: Theoretical research is conducted on the consistency issue of the maximum likelihood estimation (MLE) for the probabilistic principal component analysis (PPCA) model in the presence of rotational indeterminacy.
On the Constrained Time-Series Generation Problem
Andrea Coletta (J.P. Morgan), Svitlana Vyetrenko (J.P. Morgan)
GenerationData SynthesisOptimizationDiffusion modelTime SeriesFinance Related
🎯 What it does: This paper proposes various methods for generating time series with soft and hard constraints and validates their effectiveness in fields such as finance and energy.
On the Convergence and Sample Complexity Analysis of Deep Q-Networks with $\epsilon$-Greedy Exploration
Shuai Zhang (New Jersey Institute of Technology), Subhajit Chaudhury (IBM Research)
Convolutional Neural NetworkReinforcement LearningVideo
🎯 What it does: This paper presents the theoretical convergence and sample complexity analysis of deep Q-networks (DQN) using the ε-greedy exploration strategy for the first time, and validates the theoretical predictions through experiments.
On the Convergence of Black-Box Variational Inference
Kyurae Kim (University of Pennsylvania), Jacob R. Gardner (University of Pennsylvania)
OptimizationReinforcement LearningTabular
🎯 What it does: The paper proves the convergence of Black Box Variational Inference (BBVI) with reparameterization gradients under practical implementations and provides theoretical guarantees.
On the Convergence of CART under Sufficient Impurity Decrease Condition
Rahul Mazumder (Massachusetts Institute of Technology), Haoyue Wang (Massachusetts Institute of Technology)
🎯 What it does: The study investigates the convergence speed and error upper bound of CART decision trees in regression tasks under the condition of Sufficient Impurity Decrease (SID).
On the Convergence of Encoder-only Shallow Transformers
Yongtao Wu (Ecole Polytechnique Fédérale de Lausanne), Volkan Cevher (Ecole Polytechnique Fédérale de Lausanne)
TransformerImageText
🎯 What it does: For a single-layer shallow Transformer with an encoder, a global convergence theory for its gradient descent training is constructed and proven under achievable softmax scaling, initialization, and finite width settings, along with NTK analysis and an upper bound on the Hessian spectral norm.
On the Convergence of No-Regret Learning Dynamics in Time-Varying Games
Ioannis Anagnostides (Carnegie Mellon University), Tuomas Sandholm (Carnegie Mellon University)
OptimizationMeta Learning
🎯 What it does: This paper studies the iterative convergence properties of no-regret learning algorithms (especially optimistic gradient descent) in time-varying game environments and provides a convergence upper bound regarding the rate of game changes.
On the Convergence to a Global Solution of Shuffling-Type Gradient Algorithms
Lam M. Nguyen (IBM Research), Trang H. Tran (Cornell University)
OptimizationTabularBiomedical Data
🎯 What it does: This paper studies the properties of the shuffled version of the Stochastic Gradient Descent (SGD) algorithm in converging to a global solution under non-convex functions, proposing a new theoretical framework that demonstrates convergence in over-parameterized settings.
On the explainable properties of 1-Lipschitz Neural Networks: An Optimal Transport Perspective
Mathieu Serrurier (Université Paul-Sabatier), Thibaut Boissin (Institut de Recherche Technologique Saint-Exupéry)
Explainability and InterpretabilityAdversarial AttackConvolutional Neural NetworkImage
🎯 What it does: By training the network using optimal transport dual loss under the 1-Lipschitz constraint, this study investigates the impact of its gradient on interpretability and proposes that the Saliency Map of OTNN can serve as a high-quality explanation.
On the Exploitability of Instruction Tuning
Manli Shu (University of Maryland), Tom Goldstein (University of Maryland)
Adversarial AttackTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This paper proposes an attack method based on Automated Data Poisoning (AutoPoison), which utilizes an oracle LLM to inject malicious examples into instruction tuning data to alter the executable behavior of large language models.
On the Exploration of Local Significant Differences For Two-Sample Test
Zhijian Zhou, Wei Gao (Nanjing University)
Tabular
🎯 What it does: This paper proposes a two-sample test method called ME MaBiD based on multiple Mahalanobis kernels, and explores local significant differences through bidirectional hypothesis testing and partition trees.
On the Generalization Error of Stochastic Mirror Descent for Quadratically-Bounded Losses: an Improved Analysis
Ta Duy Nguyen (Boston University), Huy Nguyen
Optimization
🎯 What it does: This paper studies the generalization error of Stochastic Mirror Descent (SMD) under quadratically-bounded losses and proposes a new high-probability analysis method.
On the Generalization Properties of Diffusion Models
Puheng Li (Stanford University), Jiang Bian (Microsoft Research Asia)
GenerationData SynthesisDiffusion modelImageStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: Analyzed and proved the upper bound of generalization error in the training process of diffusion models, revealing the impact of early stopping strategies and mode displacement on generalization, and conducted numerical validation on synthetic and MNIST data.
On the Gini-impurity Preservation For Privacy Random Forests
XinRan Xie, Zhi-Hua Zhou (Nanjing University)
ClassificationSafty and PrivacyTabular
🎯 What it does: This paper proposes a new encryption method that maintains the key Gini impurity value during the construction process of random forests without decryption, thereby achieving privacy-preserving training and prediction of random forests.
On the Identifiability and Interpretability of Gaussian Process Models
Jiawen Chen (University of North Carolina at Chapel Hill), Didong Li (University of North Carolina at Chapel Hill)
Explainability and InterpretabilityTime Series
🎯 What it does: This paper explores and proves that in Gaussian processes (GP), when mixing Matérn kernels with different smoothness, the smoothness of the mixed kernel is determined by the least smooth component, and only the microergodic parameter of the least smooth kernel can be estimated; for multi-output GPs, it is proven that the covariance matrix in the multiplicative (separable) kernel can be uniquely determined up to a proportional constant.
On the Identifiability of Sparse ICA without Assuming Non-Gaussianity
Ignavier Ng (Carnegie Mellon University), Kun Zhang (Carnegie Mellon University)
🎯 What it does: Proposes an ICA identifiability and estimation method using second-order statistics and sparse constraints without assuming non-Gaussianity.
On the impact of activation and normalization in obtaining isometric embeddings at initialization
Amir Joudaki (ETH Zurich), Francis Bach (INRIA ENS PSL Paris)
ClassificationOptimizationImage
🎯 What it does: This study investigates how normalization layers (such as layer normalization) and nonlinear activation functions affect the isometry of the Gram matrix during the initialization of deep neural networks and quantifies its convergence rate.
On the Implicit Bias of Linear Equivariant Steerable Networks
Ziyu Chen (University of Massachusetts Amherst), Wei Zhu (University of Massachusetts Amherst)
ClassificationOptimization
🎯 What it does: This study investigates the implicit bias of gradient flow training in linear steerable networks in group-invariant binary classification tasks, and proves that the parameters of this network converge directionally to a unique maximum margin group-invariant classifier.
On the Importance of Exploration for Generalization in Reinforcement Learning
Yiding Jiang (Carnegie Mellon University), Roberta Raileanu (Meta AI Research)
Reinforcement LearningTabularBenchmark
🎯 What it does: The research explores the impact of exploration strategies on reinforcement learning generalization in Contextual MDPs (CMDP) and proposes an exploration method through Ensemble Distribution Exploration (EDE) to guide agents in exploring high-uncertainty states in the training environment.
On the Importance of Feature Separability in Predicting Out-Of-Distribution Error
RENCHUNZI XIE, Bo An (Nanyang Technological University)
ClassificationDomain AdaptationAnomaly DetectionConvolutional Neural NetworkImage
🎯 What it does: Proposes a Dispersion Score based on feature separability to estimate the model's error on out-of-distribution (OOD) data in an unlabeled setting;
On the Interplay between Social Welfare and Tractability of Equilibria
Ioannis Anagnostides (Carnegie Mellon University), Tuomas Sandholm (Carnegie Mellon University)
OptimizationReinforcement Learning
🎯 What it does: The research combines the smoothness of games with no-regret learning algorithms, proving that in games with smoothness that guarantees approximate full efficiency, Nash equilibria can be effectively approached through decentralized, computable learning dynamics (such as optimistic gradient descent); it further proposes the convergence of weak Nash equilibria in large-scale games (where the number of players is much greater than 1) and demonstrates the correspondence between smoothness and Minty properties; finally, it utilizes clairvoyant mirror descent to achieve a higher social welfare upper bound than traditional smoothness frameworks and guarantees convergence to coarse correlated equilibria (CCE).
On the Last-iterate Convergence in Time-varying Zero-sum Games: Extra Gradient Succeeds where Optimism Fails
Yi Feng (Shanghai University of Finance and Economics), Xiao Wang (Shanghai University of Finance and Economics)
Optimization
🎯 What it does: Analyzes the convergence of the last iterations of Extra Gradient (EG), Optimistic Gradient Descent Ascent (OGDA), and Negative Momentum (NM) under two types of time-varying zero-sum games (periodic and convergent perturbations).
On the Learnability of Multilabel Ranking
Vinod Raman (University of Michigan), Ambuj Tewari (University of Michigan)
ClassificationRecommendation SystemTextBiomedical Data
🎯 What it does: This paper studies the learnability of multi-label ranking problems, particularly in the context of relevance score feedback, analyzing the learning capabilities under both batch and online learning settings.
On the Minimax Regret for Online Learning with Feedback Graphs
Khaled Eldowa (Università degli Studi di Milano), Nicolò Cesa-Bianchi (Università degli Studi di Milano)
OptimizationGraph Neural NetworkReinforcement LearningGraph
🎯 What it does: This study improves the regret bounds for strongly observable undirected feedback graphs in online learning.
On the Overlooked Pitfalls of Weight Decay and How to Mitigate Them: A Gradient-Norm Perspective
Zeke Xie (University of Tokyo), Masashi Sugiyama (RIKEN Center for AIP)
OptimizationConvolutional Neural NetworkRecurrent Neural NetworkImageText
🎯 What it does: This paper systematically studies the hidden issues of weight decay from the perspective of gradient norm—non-convergence and large gradient norms—and proposes a gradient norm-based weight decay scheduler (SWD) that significantly improves the convergence speed and generalization performance of the Adam optimizer.
On the Overlooked Structure of Stochastic Gradients
Zeke Xie (Baidu Research), Ping Li (Baidu Research)
Convolutional Neural NetworkImageTabular
🎯 What it does: This paper studies the distribution and covariance structure of stochastic gradients and gradient noise in deep neural networks through formal statistical tests. It reveals that the gradient in the dimension space follows a power-law heavy tail, while the gradient in the iteration space (true gradient noise) follows a Gaussian light tail. It also finds that the covariance spectrum of stochastic gradients itself exhibits a power-law structure, further exploring its impact on the robustness of learning subspaces, batch size, label noise, network width/depth, linear networks, BatchNorm, Adam, and other factors.
On the Pareto Front of Multilingual Neural Machine Translation
Liang Chen (Peking University), Baobao Chang (Peking University)
OptimizationTransformerText
🎯 What it does: This paper studies the relationship between the performance of different language directions in multilingual neural machine translation (MNMT) and the sampling ratio, revealing the phenomenon of Pareto front collapse, and proposes a double power law model to predict the performance of each direction; based on this model, it solves the optimal sampling ratio to improve overall translation quality while maintaining training costs.
On the Planning Abilities of Large Language Models - A Critical Investigation
Karthik Valmeekam (Arizona State University), Subbarao Kambhampati (Arizona State University)
TransformerLarge Language ModelPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: Evaluate the ability of large language models in autonomous planning tasks and as heuristic tools.
On the Power of SVD in the Stochastic Block Model
Xinyu Mao (University of Southern California), Jiapeng Zhang (University of Southern California)
ClassificationOptimizationGraph
🎯 What it does: This paper studies the theoretical performance of clustering using pure SVD (without additional steps) in the Symmetric Stochastic Block Model (SSBM), proving that hidden clusters can be completely recovered over a wide range of parameters.
On the Powerfulness of Textual Outlier Exposure for Visual OoD Detection
Sangha Park (Seoul National University), Sungroh Yoon (Seoul National University)
Anomaly DetectionTransformerLarge Language ModelVision Language ModelImageTextMultimodality
🎯 What it does: The research proposes using text as proxy data for Out-of-Distribution (OoD) detection through 'Textual Outlier Exposure' and trains a linear classifier in the CLIP visual-language embedding space.
On the Properties of Kullback-Leibler Divergence Between Multivariate Gaussian Distributions
Yufeng Zhang (Hunan University), J Wang
Anomaly DetectionReinforcement LearningFlow-based ModelImage
🎯 What it does: This paper studies the Kullback–Leibler (KL) divergence between multivariate Gaussian distributions, deriving the upper and lower bounds of the reverse KL divergence when the forward KL divergence is finite, and proving that the KL divergence satisfies a relaxed triangle inequality; subsequently, these theoretical results are applied to tasks such as anomaly detection in flow models, reinforcement learning, and sample complexity.
On the Relationship Between Relevance and Conflict in Online Social Link Recommendations
Yanbang Wang (Cornell University), Jon Kleinberg (Cornell University)
Recommendation SystemOptimizationGraph Neural NetworkGraph
🎯 What it does: Analyzes the relationship between relevance and opinion conflict in social network link recommendation, providing theoretical and empirical research.
On the Robustness of Mechanism Design under Total Variation Distance
Anuran Makur (Purdue University), Athina Terzoglou (Purdue University)
🎯 What it does: This paper studies the problem of designing mechanisms under unknown and related prior distributions, particularly designing a truthful mechanism that is close to a given prior distribution in total variation distance.
On the Robustness of Removal-Based Feature Attributions
Chris Lin (University of Washington), Su-In Lee (University of Washington)
Explainability and InterpretabilityImageTabular
🎯 What it does: This paper provides a theoretical robustness analysis of feature attribution methods based on feature removal under input perturbations and model perturbations, and presents the corresponding Lipschitz upper bounds and error bounds.
On the Role of Entanglement and Statistics in Learning
Srinivasan A, Louis Schatzki (University of Illinois)
Physics Related
🎯 What it does: In the quantum statistical query (QSQ) model, the interrelationships of learning models are systematically studied: ① It is proven that for Boolean concept classes, the sample complexity using entangled measurements and that using only separable measurements are polynomially related; ② A natural concept class is constructed that exhibits exponential separation between noisy quantum PAC and QSQ learning; ③ Quantum statistical dimension (QSD) is proposed as a tool for lower bounds in QSQ learning, and it is used to provide stronger lower bounds for quantum learning tasks such as shadow projection, the Abelian hidden subgroup problem, and purity testing; ④ This work removes the previous restrictions on the diagonalizability of measurements.
On the Role of Noise in the Sample Complexity of Learning Recurrent Neural Networks: Exponential Gaps for Long Sequences
Alireza Fathollah Pour, Hassan Ashtiani (McMaster University)
ClassificationRecurrent Neural NetworkSequential
🎯 What it does: This paper studies the PAC learning sample complexity of multi-layer Sigmoid RNNs with Gaussian noise in sequence classification tasks, providing both upper and lower bounds.
On the Role of Randomization in Adversarially Robust Classification
Lucas Gnecco Heredia (CNRS LAMSADE Université Paris Dauphine PSL), Yann Chevaleyre (CNRS LAMSADE Université Paris Dauphine PSL)
ClassificationAdversarial Attack
🎯 What it does: This paper theoretically analyzes the role of randomization in adversarially robust classification, exploring whether randomized ensembles (REC) and noise injection can outperform the optimal deterministic classifiers in terms of adversarial risk when given a benchmark set of deterministic classifiers. It proves that any binary probabilistic classifier can be mapped to one or more deterministic threshold classifiers with at least equivalent robustness; it also points out that randomization does not provide additional advantages when the closure property of the set is satisfied.
On the Size and Approximation Error of Distilled Datasets
Alaa Maalouf (Massachusetts Institute of Technology), Daniela Rus (Massachusetts Institute of Technology)
Knowledge DistillationImageTabular
🎯 What it does: This study explores dataset distillation from a theoretical perspective, providing bounds on the size of the distilled dataset and the approximation error, particularly under the kernel ridge regression (KRR) setting, proving the existence of small distilled datasets and their corresponding excess risk.
On the spectral bias of two-layer linear networks
Aditya Vardhan Varre (École Polytechnique Fédérale de Lausanne), Nicolas Flammarion (École Polytechnique Fédérale de Lausanne)
OptimizationRepresentation LearningTabular
🎯 What it does: This paper studies the behavior of a two-layer fully connected linear network trained using gradient flow under squared loss, revealing the relationship between the implicit bias of the optimization process and the scale of initialization.
On the Stability-Plasticity Dilemma in Continual Meta-Learning: Theory and Algorithm
Qi CHEN, Mario Marchand (Laval University)
Meta LearningTabular
🎯 What it does: This paper proposes a theory-driven framework for Continual Meta-Learning (CML) to address the stability-plasticity dilemma, and presents the corresponding Average Excess Risk (AER) objective and the dynamic scheduling algorithm DCML.
On the Statistical Consistency of Risk-Sensitive Bayesian Decision-Making
Prateek Jaiswal (Texas A&M University), Vinayak Rao
Tabular
🎯 What it does: This paper proposes a risk-sensitive variational Bayesian (RSVB) framework for simultaneously approximating the posterior and solving decision-making problems in cases where the posterior is intractable;
On the Sublinear Regret of GP-UCB
Justin Whitehouse (Carnegie Mellon University), Steven Wu
🎯 What it does: This paper improves the tuning and lower bounds of the Gaussian Process Upper Confidence Bound (GP-UCB) algorithm in the kernelized multi-armed bandit problem, proving that GP-UCB can achieve sublinear cumulative regret when the kernel function satisfies polynomial eigenvalue decay (especially for the Matérn kernel).
On the Trade-off of Intra-/Inter-class Diversity for Supervised Pre-training
Jieyu Zhang (University of Washington), Alexander Ratner (University of Washington)
ClassificationRepresentation LearningConvolutional Neural NetworkSupervised Fine-TuningImage
🎯 What it does: This paper studies the impact of the trade-off between intra-class diversity (number of samples per class) and inter-class diversity (number of classes) on downstream task performance using a fixed-size supervised pre-training dataset, proposing and validating the optimal class-to-sample ratio.
On the Variance, Admissibility, and Stability of Empirical Risk Minimization
Gil Kur, Alexander Rakhlin (Massachusetts Institute of Technology)
Optimization
🎯 What it does: This paper studies the variance, acceptability, and stability of empirical risk minimization (ERM) under fixed and random designs. The authors prove that in the class of generalized (closed convex) functions, the variance term of ERM matches the lower bound, indicating that if the overall performance of ERM is suboptimal, the reason can only stem from the bias term. Furthermore, the authors provide a non-asymptotic upper bound on variance in random design based on isometry remainders and Lipschitz Concentration Property (LCP), and they reprove Chatterjee's acceptability theorem, extending it to random design. Finally, the authors analyze the stability of approximate minimizers of ERM and the irregularity of the loss landscape in non-Donsker classes.
On Transfer of Adversarial Robustness from Pretraining to Downstream Tasks
Laura Fee Nern, Yash Sharma (University of Tübingen)
Representation LearningAdversarial AttackConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: This study proves that the representation layer of pre-trained models constrains the adversarial robustness of downstream linear predictors, providing corresponding theoretical upper bounds and criteria.
On-the-Fly Adapting Code Summarization on Trainable Cost-Effective Language Models
Yufan Cai (Shanghai Jiao Tong University National University of Singapore), Jin Song Dong (National University of Singapore)
GenerationAI Code AssistantTransformerSupervised Fine-TuningText
🎯 What it does: Introducing AdaCom, an online adaptive framework that fine-tunes the code comment generation model locally during inference using a small number of training samples, thereby improving the model's performance on different project codes.
One Fits All: Power General Time Series Analysis by Pretrained LM
Tian Zhou (Alibaba Inc.), Rong Jin (Alibaba Inc.)
ClassificationAnomaly DetectionTransformerLarge Language ModelSupervised Fine-TuningTime Series
🎯 What it does: This paper proposes a general time series analysis framework—Frozen Pretrained Transformer (FPT). It transfers Transformers (such as GPT-2, BERT, BEiT) that have been pretrained on large-scale text or image data to time series tasks, requiring only the redesign of input embeddings, instance normalization, and output layers, and fine-tuning these layers while keeping other self-attention and feedforward layers frozen. The framework is evaluated on seven major tasks: classification, anomaly detection, imputation, short-term/long-term forecasting, and few-shot/zero-shot prediction.
One Less Reason for Filter Pruning: Gaining Free Adversarial Robustness with Structured Grouped Kernel Pruning
Shaochen Zhong (Rice University), Xia Hu (Rice University)
CompressionAdversarial AttackConvolutional Neural NetworkImage
🎯 What it does: A pruning method for grouped convolution kernels called SR-GKP is proposed and implemented, aimed at enhancing the adversarial robustness of structured pruning models.
One Risk to Rule Them All: A Risk-Sensitive Perspective on Model-Based Offline Reinforcement Learning
Marc Rigter (Oxford Robotics Institute University of Oxford), Nick Hawes (Oxford Robotics Institute University of Oxford)
Reinforcement LearningTabular
🎯 What it does: A model-based offline reinforcement learning algorithm named 1R2R is proposed, which utilizes risk sensitivity to simultaneously avoid uncertainties caused by distribution shift and environmental randomness;
One-2-3-45: Any Single Image to 3D Mesh in 45 Seconds without Per-Shape Optimization
Minghua Liu (University of California San Diego), Hao Su (University of California San Diego)
GenerationDepth EstimationDiffusion modelImageMesh
🎯 What it does: Proposes an end-to-end method for quickly generating high-quality 360° 3D meshes from a single image.
One-for-All: Bridge the Gap Between Heterogeneous Architectures in Knowledge Distillation
Zhiwei Hao (Beijing Institute of Technology), Chang Xu (University of Sydney)
Knowledge DistillationConvolutional Neural NetworkTransformerImage
🎯 What it does: This paper proposes a unified cross-architecture knowledge distillation framework, OFA-KD, which allows different structured teacher models (CNN, Transformer, MLP) to effectively distill into a student model.
One-Line-of-Code Data Mollification Improves Optimization of Likelihood-based Generative Models
Ba-Hien Tran (EURECOM), Maurizio Filippone (EURECOM)
GenerationOptimizationFlow-based ModelAuto EncoderImageTabular
🎯 What it does: A method for continuous optimization in likelihood-based generative model training is proposed, which achieves this through progressive denoising (Gaussian or blurred), allowing for improved generation quality and density estimation with just a single line of code.
One-Pass Distribution Sketch for Measuring Data Heterogeneity in Federated Learning
Zichang Liu (Rice University), Anshumali Shrivastava (ThirdAI Corporation)
Federated LearningImageText
🎯 What it does: A one-pass distribution sketch is proposed to quantify data heterogeneity in federated learning, and this sketch is used for client selection and cold start model retrieval.
One-step differentiation of iterative algorithms
Jerome Bolte, Samuel Vaiter (CNRS and Université Côte d'Azur)
OptimizationAuto EncoderTabular
🎯 What it does: The study and theoretical analysis focus on the Jacobian-free backpropagation method for iterative algorithms, providing a quantitative relationship between error and convergence rate.
One-Step Diffusion Distillation via Deep Equilibrium Models
Zhengyang Geng (Carnegie Mellon University), J Zico Kolter
GenerationKnowledge DistillationTransformerDiffusion modelImage
🎯 What it does: This paper proposes an offline distillation method that compresses the multi-step sampling process of diffusion models into a single-step generative model, and designs the Generative Equilibrium Transformer (GET) architecture to achieve one-shot sampling.
OneNet: Enhancing Time Series Forecasting Models under Concept Drift by Online Ensembling
YiFan Zhang, Tieniu Tan (University of Chinese Academy of Sciences)
Reinforcement LearningTime Series
🎯 What it does: A dual-stream online model called OneNet is proposed, which uses two predictors for cross-time and cross-variable, and dynamically integrates their outputs through online convex programming (OCP) to achieve time series forecasting in the context of concept drift.
Online (Multinomial) Logistic Bandit: Improved Regret and Constant Computation Cost
Yu-Jie Zhang (University of Tokyo), Masashi Sugiyama (RIKEN AIP)
Recommendation SystemOptimizationComputational EfficiencyReinforcement LearningTabular
🎯 What it does: A polynomial logic Bandit algorithm OFUL-MLogB is proposed, which takes into account both statistical and computational efficiency, supporting both binary and multi-valued feedback scenarios.
Online Ad Allocation with Predictions
Fabian Christian Spaeh (Boston University), Alina Ene (Boston University)
OptimizationTabular
🎯 What it does: A prediction-based online advertising allocation algorithm is proposed, suitable for Display Ads and the Generalized Assignment Problem (GAP), achieving instant allocation through the introduction of a free disposal model under budget constraints.
Online Ad Procurement in Non-stationary Autobidding Worlds
Jason Cheuk Nam Liang (Massachusetts Institute of Technology), Baoyu Zhou (University of Michigan)
OptimizationReinforcement Learning
🎯 What it does: An online learning framework is proposed to help advertisers dynamically optimize leverage decisions on advertising platforms in a non-stationary automated bidding environment while satisfying long-term constraints.
Online Adaptive Policy Selection in Time-Varying Systems: No-Regret via Contractive Perturbations
Yiheng Lin (California Institute of Technology), Adam Wierman (California Institute of Technology)
OptimizationReinforcement Learning
🎯 What it does: This study investigates the problem of online adaptive strategy selection in dynamically changing systems, proposing a gradient-based adaptive strategy selection algorithm (GAPS) and establishing a general online optimization analysis framework.
Online Clustering of Bandits with Misspecified User Models
Zhiyong Wang (Chinese University of Hong Kong), John C.S. Lui (Chinese University of Hong Kong)
Recommendation SystemOptimizationReinforcement LearningTabular
🎯 What it does: This paper presents the Bandits clustering problem under user model misspecification and introduces the CBMUM framework; it also designs robust clustering algorithms RCLUMB and RSCLUMB, which can efficiently perform user grouping and recommendation in misspecified models.
Online Constrained Meta-Learning: Provable Guarantees for Generalization
Siyuan Xu (Pennsylvania State University), Minghui Zhu (Pennsylvania State University)
OptimizationMeta LearningImage
🎯 What it does: An online constrained meta-learning framework is proposed, which can continuously learn meta-knowledge from constrained sequential learning tasks and provides upper bounds on the optimality gap of task-specific models and constraint violations.
Online Control for Meta-optimization
Xinyi Chen (Princeton University), Elad Hazan (Princeton University)
OptimizationMeta LearningTabular
🎯 What it does: This paper proposes a framework based on online non-stochastic control to learn the optimal optimizer (i.e., meta-optimization) for different optimization tasks. By using control theory, the meta-optimization problem is transformed into an optimal control problem, and a controller is designed to achieve algorithm adaptability.
Online Convex Optimization with Unbounded Memory
Raunak Kumar (Cornell University), Robert Kleinberg (Cornell University)
Optimization
🎯 What it does: A new framework for online convex optimization is proposed, allowing the current round's loss to depend on the history of all past decisions.
Online Corrupted User Detection and Regret Minimization
Zhiyong Wang (Chinese University of Hong Kong), John C.S. Lui (Chinese University of Hong Kong)
Recommendation SystemAnomaly DetectionOptimizationReinforcement LearningTabular
🎯 What it does: This paper proposes the problem of Online User Relationship Learning and Damaged User Detection (LOCUD), and designs a robust clustering Bandit algorithm RCLUB-WCU and an online detection algorithm OCCUD, which can effectively learn user preferences and clustering in the presence of damaged users and identify damaged users online.